I think there was a lot more brainpower invested in the media generation side of things. The noise-based diffusion technique is further developed. It had a dis

I think there was a lot more brainpower invested in the media generation side of things. The noise-based diffusion technique is further developed. It had a discovery of applying a physics-based understanding of Brownian motion to guide it. Image generation has comparatively simple training process - this is an image with dog, and without dog (contrastive learning). Might be worth to watch the diffusion based LLMs.